When Should a Failing Robot Ask? Initiating Corrective Human-Robot Dialogue from Audited Sensor Evidence
cs.RO · cs.AI · cs.HC
Quick verdict
Claims frame-only VLMs never beat the majority baseline on robot-failure diagnosis and ignore ask costs; rigorous leakage-audited simulation study, convincing.
Summary
This paper builds a simulated benchmark (LIBERO pick-and-place) in which manipulation failures carry injected, known causes, then audits what each sensor reveals and tests whether vision-language models can use that evidence to decide when to act, consult their own sensors, or ask a human. A leakage audit (shuffle, transfer, separation checks) shows detection failures are diagnosable from images (0.910, surviving held-out viewpoints); grasp failures are diagnosed almost perfectly from force-torque data (0.986) but not from images (classifier plateau 0.545); and the placement family's 0.998 image score collapses to 0.605 under held-out camera angles, disqualifying it as a shortcut. Six open 7B–16B models score at or below the majority-class baseline from frames under every prompt variant; refusing versus answering tracks option position rather than evidence (moving the refusal option from last to first collapses refusal from 78–100% to 0–6% in three of six swept pairs), and stated confidence is uninformative. Serializing force data as ten lines of text yields the first above-baseline diagnoses in four of six models (0.457–0.543 vs 0.400). Against a cost-optimal act/ask/sense policy, no model's ask rate tracks a fourfold question-cost sweep, though one human answer lifts accuracy to 0.70–0.81. Limitations: simulation only, scripted policy and answerer, small test sets (35–36 episodes), stipulated costs, no frontier models.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Diagnosability of manipulation failures is sensor- and family-specific: detection failures reach 0.910 from images, grasp failures 0.986 from force data but only a 0.545 plateau from images, and the placement family's 0.998 image score is a viewpoint shortcut that fails the transfer check.
- Six open 7B–16B vision-language models never exceed the majority-class baseline from frames under any prompt variant, and their refusal versus commitment is governed by option position rather than evidence.
- Model confidence — stated ratings, token probabilities, and sample frequency — carries essentially no information about correctness, with Pixtral's token log-probability the sole exception (ρ=+0.50, recovering 56% of the oracle selective-asking value).
- Serializing force data as ten lines of text yields the first above-baseline diagnoses in four of six models (0.457–0.543 vs 0.400), showing the frames-only failure largely reflects missing sensor input rather than missing ability.
- No model's ask rate responds to a fourfold change in question cost, yet a single menu-phrased human answer lifts final accuracy to 0.70–0.81, close to the scripted answerer's own 0.80 reliability.
Strengths
Leakage audit is a genuine methodological contribution for failure benchmarks
Three checks (shuffle, transfer, separation; Sec. III-B/IV-B) plus severity-band transfer. The placement family shows the audit biting: a 0.998 image score collapses to 0.605 under held-out azimuth (1.000→0.683 under contrast normalization) and is disqualified as a container-position shortcut (Table I, App. A), while detection's 0.910 survives (0.844/0.813 held-out) — the checks discriminate, they are not ritual.
Unusually disciplined negative-result reporting
The forced-answer round was pre-registered in writing and reported alongside, never instead (Sec. V); three confidence readouts are separated; refused episodes' token probabilities show the refusals hid nothing (largest difference 0.07 vs forced-answer accuracy); the construction confound (unrendered grasp physics vs rendered detection causes) is disclosed in the Introduction.
Actionable central finding on sensor routing
Decision-theoretic scoring instead of qualitative verdicts
Weaknesses
External validity confined to one simulator and a scripted policy
Thin samples underpin the decision layer
Stipulated costs and a scripted, wording-insensitive answerer
No code or data release stated anywhere
Model scope excludes frontier systems and any fine-tuning
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper studies the first move of corrective human-robot dialogue: after a manipulation failure, should the robot act on its own diagnosis, consult an onboard sensor, or interrupt a human? Three claims are chained: (1) an audited measurement of what each sensor reveals per failure family (Sec. IV-B, Table I); (2) evidence that six open 7B–16B vision-language models diagnose at or below a majority-class baseline from frames, with refusal versus commitment governed by option position rather than evidence (Sec. VI-B, Tables II and VII); (3) a cost-optimal act/sense/ask policy against which models are scored as regret (Sec. III-C, VI-D, Table VIII). Scoping is exemplary for this genre: negative sensor claims are stated as classifier-sequence plateaus, never ceilings, and the construction confound — grasp causes are physics the renderer never draws, detection causes are rendered scene properties — is disclosed in the Introduction rather than buried.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the paper 'When Should a Failing Robot Ask?' about?
It studies the moment right after a robot fails a task: should it act on its own diagnosis, check another onboard sensor, or interrupt a human with a question? The authors built a simulation benchmark where failure causes are injected and therefore known, measured what camera images versus force sensors reveal about each failure type, and found that current open vision-language models diagnose poorly from images and decide badly about when to ask.
Can a robot diagnose its own grasp failures from camera images?
In this study, essentially no. Grasp failures are caused by physical parameters (grip force, object mass, surface friction) that a camera cannot see; a sequence of increasingly capable image classifiers plateaued at 0.545 accuracy. The robot's own wrist force-torque and gripper data diagnose the same failures at 0.986.
Did the AI models improve when given the robot's sensor data?
Yes, partly. Converting force data into ten lines of text let four of the six tested models beat the majority-class baseline for the first time (0.457–0.543 vs 0.400), though still far below the 0.986 audit score. One model, Cosmos-Reason2-8B, could not use the telemetry in any form.
When should a robot ask a human for help instead of retrying?
The paper frames it as a cost calculation: with a wrong repair costing ten times a retry and a question costing three retries, asking pays when the robot's own diagnosis accuracy is below roughly 0.59. The tested models ignored question costs entirely, but a single human answer raised their final accuracy to 0.70–0.81, close to the simulated helper's own 0.80 reliability.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–4 are available only as extracted captions and axis labels; plotted curves could not be visually inspected.
- Appendices A–H appear fully extracted but table formatting is lossy; numbers were cross-checked between text and tables where duplicated.
- arXiv preprint (Sep 2026), not peer reviewed at the time of this review.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.